Research On Super-resolution Reconstruction And Recognition Algorithm Of Motion Blurred License Plate Based On Deep Learning

License plate as the most direct identification of motor vehicles,in traffic management,security monitoring,intelligent parking and other fields have a wide range of applications.However,affected by factors such as shaking of filming equipment,rain and snow,and filming distance,the taken vehicle information is usually of poor quality and the license plate characters are blurred,and the existing license plate recognition algorithms are not ideal for recognition in complex scenes.This paper conducts research on the above motion blur license plate reconstruction and recognition technology and integrates the corresponding research results into an intelligent license plate reconstruction and recognition system.The main research contents and innovation points are as follows:(1)To address the shortcomings of ESRGAN in the traditional convolutional approach,this paper proposes a multi-scale convolutional and reconstruction network incorporating the attention mechanism.Firstly,the feature extraction module of the original model is improved,and Res2Net is introduced inside the RRDB module to increase the perceptual field of the network and fully learn the rich multiscale information.Secondly,the ECA module is connected to the tail of RRDB to learn the weight information between different channels and improve the overall reconstruction ability of the model.Inspired by the MSB module,the network finally introduces a stacked multiscale convolution module to enhance the nonlinear representation of the upsampling process.In the motion blur license plate dataset produced in this paper,the proposed algorithm has some improvement over the comparison model in both subjective and objective evaluation metrics.(2)For license plate character recognition in complex scenes,this paper compares and analyzes two deep learning-based license plate character recognition models,CRNN and LPR-Net,and proposes an improved CRNN network by combining the advantages of both algorithms.For the tilted license plate scene,the STN network is firstly introduced to pre-correct the tilted license plate image.Then,the depth-separable convolution is used to replace part of the traditional convolution of CNN layers to reduce the processing time of the network.Some pooling windows are modified to 1×2 for the license plate size,and a BN layer is added at the end of the CNN to accelerate the convergence of the network.The recurrent layer consists of a deep bidirectional GRU structure.Finally,the CTC layer is transcribed to predict the character text of the license plate image.The improved CRNN network significantly improves the recognition accuracy of license plate characters at the cost of a small increase in processing time.(3)Based on the above research results and image processing techniques,an intelligent license plate reconstruction and recognition system is designed and implemented in this paper.The system mainly includes three modules: image processing,license plate reconstruction and license plate recognition.The image processing module mainly provides visualization pages for image input,license plate detection and license plate correction functions;the license plate reconstruction module mainly provides visualization pages for image input and super-resolution reconstruction;the license plate recognition module mainly provides visualization pages for image input and character recognition.